Pizza Bot

Overview

Pizza Bot is a self-hosted desktop application that gives AI agents an inbox — a mailbox where background AI agents deliver their outputs to you, rather than you running agents synchronously and waiting for results in a terminal. The tool launched on Hacker News on September 15, 2026 (59 points), and was open-sourced under Apache 2.0 after internal use by 2,000+ people at Amazon.

The pitch is direct: "We'd like to do for knowledge workers what Claude Code and Codex have done for programmers." Pizza Bot targets the class of work that doesn't fit in a terminal — long-running research, scheduled data pulls, draft summaries, background automation — and packages it as a personal notification inbox.

The application is built on DeepAgents and LangGraph for state management, with MCP servers as the tool layer. Provider choice is deliberately open: Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, or local models via Ollama. Available for Mac, Windows, and Linux.

Features

• Background agent inbox: Agents run in the background and deliver outputs to a personal email-style inbox
• Self-hosted: Runs entirely on your machine; no cloud lock-in
• Provider-agnostic: Works with Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, or Ollama for local models
• Built on modern primitives: DeepAgents + LangGraph for state management, MCP servers as tool layer
• Cross-platform: Mac, Windows, Linux
• Apache 2.0 license: Fully open source
• Enterprise-proven: 2,000+ internal users at Amazon before open-sourcing

Installation

# macOS (Homebrew)
brew install pizza-bot/tap/pizza-bot

# Or via Docker
docker pull pizza-bot-app/pizza-bot
docker run -p 8765:8765 pizza-bot-app/pizza-bot

# Linux
curl -fsSL https://get.pizza-bot.app | sh

Configure providers via the built-in settings UI, or programmatically:

# pizza-bot.yaml
providers:
  default: anthropic
  fallback: openrouter
  local:
    provider: ollama
    models: [llama3.1, qwen2.5]

Quick Start

# Define an agent and submit a background task
from pizza_bot import Agent, schedule

agent = Agent(
    model="claude-opus-4-7",
    tools=["mcp://github", "mcp://notion"],
    system_prompt="You research technical topics and produce concise briefs."
)

# One-shot: result lands in your inbox when done
agent.send("Summarize the top 5 Rust MCP servers this month")

# Scheduled: runs daily at 09:00
schedule(agent, "cron", "0 9 * * *", "Compile yesterday's HN top AI stories")

Architecture

┌─────────────────────────────────────────┐
│           Pizza Bot Desktop             │
│  ┌──────────────────────────────────┐   │
│  │  Inbox (delivery + threading)    │   │
│  ├──────────────────────────────────┤   │
│  │  Agent Runner (LangGraph state)  │   │
│  ├──────────────────────────────────┤   │
│  │  DeepAgents orchestration layer  │   │
│  ├──────────────────────────────────┤   │
│  │  MCP tool client                 │   │
│  └──────────────────────────────────┘   │
└─────────────────────────────────────────┘
              │           │
              ▼           ▼
        LLM Providers   MCP Servers
   (Anthropic/OpenAI/  (filesystem, DB,
    Gemini/Ollama/       GitHub, etc.)
    Bedrock)

Use Cases

• Morning briefings: Scheduled agents compile news, calendar, and metrics each morning
• Async research: Submit a research task before lunch; read the report at dinner
• Data pulls: Background agents fetch reports from internal tools on a cron schedule
• Drafting pipeline: Generate first drafts of recurring documents while you do other work

Pros

• ✅ Agents run in the background; not bound to a terminal
• ✅ Delivery model is intuitive (inbox, not polling)
• ✅ Provider-agnostic; switch models freely
• ✅ Enterprise-proven at Amazon before open-sourcing
• ✅ Built on modern primitives (MCP, LangGraph, DeepAgents)
• ✅ Self-hosted; data stays local

Cons

• ❌ Desktop-only delivery; no mobile app yet
• ❌ Requires configuring MCP servers and providers up-front
• ❌ Smaller community than general-purpose agent frameworks
• ❌ Inbox threading is opinionated; some users prefer flat lists

When to Use

• Use when: You have background AI tasks (research, summarization, scheduled pulls) that don't fit in a terminal
• Avoid when: You need agents that respond interactively in real time, or if you want a no-config demo tool

Resources

• GitHub Repository
• Pizza Bot Documentation
• DeepAgents (state management)
• LangGraph